feat(jetson): Support for Jetpack 4.6

This commit adds support for Jetpack 4.6. Users should now add the
--platforms flag to bazel compilation to target the jetpack version

e.g. `bazel build //:libtrtorch --platforms //toolchains:jetpack_4.6`

By default setup.py now expects Jetpack 4.6. To override add the
`--jetpack-version 4.5` flag

Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>

Author: narendasan

core/conversion/conversionctx/ConversionCtx.h

@@ -24,7 +24,7 @@ struct Device {
 };
 
 struct BuilderSettings {
-  std::set<nvinfer1::DataType> enabled_precisions = {nvinfer1::DataType::kFLOAT};
+  std::set<nvinfer1::DataType> enabled_precisions = {};
   bool sparse_weights = false;
   bool disable_tf32 = false;
   bool refit = false;

docsrc/tutorials/installation.rst

@@ -237,7 +237,7 @@ Install or compile a build of PyTorch/LibTorch for aarch64
 
 NVIDIA hosts builds the latest release branch for Jetson here:
 
-    https://forums.developer.nvidia.com/t/pytorch-for-jetson-nano-version-1-5-0-now-available/72048
+    https://forums.developer.nvidia.com/t/pytorch-for-jetson-version-1-9-0-now-available/72048
 
 
 Enviorment Setup
@@ -285,29 +285,10 @@ To build natively on aarch64-linux-gnu platform, configure the ``WORKSPACE`` wit
     #    strip_prefix = "TensorRT-7.1.3.4"
     #)
 
+    NOTE: You may also need to configure the CUDA version to 10.2 by setting the path for the cuda new_local_repository
 
-2. Disable Python API testing dependencies:
-
-.. code-block:: shell
-
-    #pip3_import(
-    #    name = "trtorch_py_deps",
-    #    requirements = "//py:requirements.txt"
-    #)
-
-    #load("@trtorch_py_deps//:requirements.bzl", "pip_install")
-    #pip_install()
-
-    #pip3_import(
-    #   name = "py_test_deps",
-    #   requirements = "//tests/py:requirements.txt"
-    #)
-
-    #load("@py_test_deps//:requirements.bzl", "pip_install")
-    #pip_install()
 
-
-3. Configure the correct paths to directory roots containing local dependencies in the ``new_local_repository`` rules:
+2. Configure the correct paths to directory roots containing local dependencies in the ``new_local_repository`` rules:
 
     NOTE: If you installed PyTorch using a pip package, the correct path is the path to the root of the python torch package.
     In the case that you installed with ``sudo pip install`` this will be ``/usr/local/lib/python3.6/dist-packages/torch``.
@@ -346,19 +327,30 @@ use that library, set the paths to the same path but when you compile make sure
 Compile C++ Library and Compiler CLI
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 
+    NOTE: Due to shifting dependency locations between Jetpack 4.5 and 4.6 there is a now a flag to inform bazel of the Jetpack version 
+
+    .. code-block:: shell
+
+        --platforms //toolchains:jetpack_4.x
+
+
 Compile TRTorch library using bazel command:
 
 .. code-block:: shell
 
-   bazel build //:libtrtorch
+   bazel build //:libtrtorch --platforms //toolchains:jetpack_4.6
 
 Compile Python API
 ^^^^^^^^^^^^^^^^^^^^
 
+    NOTE: Due to shifting dependencies locations between Jetpack 4.5 and Jetpack 4.6 there is now a flag for ``setup.py`` which sets the jetpack version (default: 4.6)
+
 Compile the Python API using the following command from the ``//py`` directory:
 
 .. code-block:: shell
 
     python3 setup.py install --use-cxx11-abi
 
-If you have a build of PyTorch that uses Pre-CXX11 ABI drop the ``--use-cxx11-abi`` flag
+If you have a build of PyTorch that uses Pre-CXX11 ABI drop the ``--use-cxx11-abi`` flag
+
+If you are building for Jetpack 4.5 add the ``--jetpack-version 4.5`` flag

py/setup.py

@@ -13,17 +13,35 @@
 from shutil import copyfile, rmtree
 
 import subprocess
+import platform
+import warnings
 
 dir_path = os.path.dirname(os.path.realpath(__file__))
 
 __version__ = '0.4.0a0'
 
 CXX11_ABI = False
 
+JETPACK_VERSION = None
+
 if "--use-cxx11-abi" in sys.argv:
     sys.argv.remove("--use-cxx11-abi")
     CXX11_ABI = True
 
+if platform.uname().processor == "aarch64":
+    if "--jetpack-version" in sys.argv:
+        version_idx = sys.argv.index("--jetpack-version") + 1
+        version = sys.argv[version_idx]
+        sys.argv.remove(version)
+        sys.argv.remove("--jetpack-version")
+        if version == "4.5":
+            JETPACK_VERSION = "4.5"
+        elif version == "4.6":
+            JETPACK_VERSION = "4.6"
+    if not JETPACK_VERSION:
+        warnings.warn("Assuming jetpack version to be 4.6, if not use the --jetpack-version option")
+        JETPACK_VERSION = "4.6"
+
 
 def which(program):
     import os
@@ -66,6 +84,13 @@ def build_libtrtorch_pre_cxx11_abi(develop=True, use_dist_dir=True, cxx11_abi=Fa
     else:
         print("using CXX11 ABI build")
 
+    if JETPACK_VERSION == "4.5":
+        cmd.append("--platforms=//toolchains:jetpack_4.5")
+        print("Jetpack version: 4.5")
+    elif JETPACK_VERSION == "4.6":
+        cmd.append("--platforms=//toolchains:jetpack_4.6")
+        print("Jetpack version: 4.6")
+
     print("building libtrtorch")
     status_code = subprocess.run(cmd).returncode
 

py/trtorch/csrc/tensorrt_classes.h

@@ -155,7 +155,7 @@ struct CompileSpec : torch::CustomClassHolder {
 
   std::vector<Input> inputs;
   nvinfer1::IInt8Calibrator* ptq_calibrator = nullptr;
-  std::set<DataType> enabled_precisions = {DataType::kFloat};
+  std::set<DataType> enabled_precisions = {};
   bool sparse_weights = false;
   bool disable_tf32 = false;
   bool refit = false;

tests/core/conversion/evaluators/test_aten_evaluators.cpp

@@ -3,6 +3,7 @@
 #include "gtest/gtest.h"
 #include "tests/util/util.h"
 #include "torch/csrc/jit/ir/irparser.h"
+#include "torch/torch.h"
 
 TEST(Evaluators, DivIntEvaluatesCorrectly) {
   const auto graph = R"IR(

tests/cpp/BUILD

@@ -77,6 +77,7 @@ cc_test(
     deps = [
         ":cpp_api_test",
     ],
+    timeout="long"
 )
 
 cc_test(

tests/modules/hub.py

@@ -4,6 +4,8 @@
 import torchvision.models as models
 import timm
 
+torch.hub._validate_not_a_forked_repo = lambda a, b, c: True
+
 models = {
     "alexnet": {
         "model": models.alexnet(pretrained=True),

tests/modules/requirements.txt

@@ -0,0 +1 @@
+timm==v0.4.12

tests/py/test_api_dla.py

@@ -37,7 +37,7 @@ def test_compile_traced(self):
                 "dla_core": 0,
                 "allow_gpu_fallback": True
             },
-            "enabled_precision": {torch.float, torch.half}
+            "enabled_precisions": {torch.half}
         }
 
         trt_mod = trtorch.compile(self.traced_model, compile_spec)
@@ -53,7 +53,7 @@ def test_compile_script(self):
                 "dla_core": 0,
                 "allow_gpu_fallback": True
             },
-            "enabled_precision": {torch.float, torch.half}
+            "enabled_precisions": {torch.half}
         }
 
         trt_mod = trtorch.compile(self.scripted_model, compile_spec)

third_party/cublas/BUILD

@@ -3,11 +3,21 @@ package(default_visibility = ["//visibility:public"])
 # NOTE: This BUILD file is only really targeted at aarch64, the rest of the configuration is just to satisfy bazel, x86 uses the cublas source from the CUDA build file since it will be versioned with CUDA.
 
 config_setting(
-    name = "aarch64_linux",
+    name = "jetpack_4.5",
     constraint_values = [
         "@platforms//cpu:aarch64",
         "@platforms//os:linux",
-    ],
+        "@//toolchains/jetpack:4.5"
+    ]
+)
+
+config_setting(
+    name = "jetpack_4.6",
+    constraint_values = [
+        "@platforms//cpu:aarch64",
+        "@platforms//os:linux",
+        "@//toolchains/jetpack:4.6"
+    ]
 )
 
 config_setting(
@@ -20,7 +30,8 @@ config_setting(
 cc_library(
     name = "cublas_headers",
     hdrs = select({
-        ":aarch64_linux": ["include/cublas.h"] + glob(["usr/include/cublas+.h"]),
+        ":jetpack_4.5": ["include/cublas.h"] + glob(["usr/include/cublas+.h"]),
+        ":jetpack_4.6": ["local/cuda/include/cublas.h"] + glob(["usr/cuda/include/cublas+.h"]),
         "//conditions:default": ["local/cuda/include/cublas.h"] + glob(["usr/cuda/include/cublas+.h"]),
     }),
     includes = ["include/"],
@@ -30,7 +41,8 @@ cc_library(
 cc_import(
     name = "cublas_lib",
     shared_library = select({
-        ":aarch64_linux": "lib/aarch64-linux-gnu/libcublas.so",
+        ":jetpack_4.5": "lib/aarch64-linux-gnu/libcublas.so",
+        ":jetpack_4.6": "local/cuda/targets/aarch64-linux/lib/libcublas.so",
         ":windows": "lib/x64/cublas.lib",
         "//conditions:default": "local/cuda/targets/x86_64-linux/lib/libcublas.so",
     }),
@@ -40,7 +52,8 @@ cc_import(
 cc_import(
     name = "cublas_lt_lib",
     shared_library = select({
-        ":aarch64_linux": "lib/aarch64-linux-gnu/libcublasLt.so",
+        ":jetpack_4.5": "lib/aarch64-linux-gnu/libcublasLt.so",
+        ":jetpack_4.6": "local/cuda/targets/aarch64-linux/lib/libcublasLt.so",
         "//conditions:default": "local/cuda/targets/x86_64-linux/lib/libcublasLt.so",
     }),
     visibility = ["//visibility:private"],

third_party/tensorrt/local/BUILD

@@ -319,5 +319,8 @@ cc_library(
     ],
     linkopts = [
         "-lpthread",
-    ]
+    ] + select({
+        ":aarch64_linux": ["-Wl,--no-as-needed -ldl -lrt -Wl,--as-needed"],
+        "//conditions:default": []
+    })
 )

toolchains/BUILD

@@ -7,3 +7,21 @@ platform(
         "@platforms//cpu:aarch64",
     ],
 )
+
+platform(
+    name = "jetpack_4.5",
+    constraint_values = [
+        "@platforms//os:linux",
+        "@platforms//cpu:aarch64",
+        "@//toolchains/jetpack:4.5"
+    ]
+)
+
+platform(
+    name = "jetpack_4.6",
+    constraint_values = [
+        "@platforms//os:linux",
+        "@platforms//cpu:aarch64",
+        "@//toolchains/jetpack:4.6"
+    ]
+)

toolchains/jetpack/BUILD

@@ -0,0 +1,11 @@
+package(default_visibility = ["//visibility:public"])
+
+constraint_setting(name = "jetpack")
+constraint_value(
+    name = "4.5",
+    constraint_setting = ":jetpack"
+)
+constraint_value(
+    name = "4.6",
+    constraint_setting = ":jetpack"
+)